On the Robustness of Explanations of Deep Neural Network Models: A Survey

Explanations of AI models can be manipulated by small input changes: The survey reveals that explanations generated by deep neural networks—which are supposed to show why a model made a decision—can be distorted through minor perturbations to the input data, raising concerns about whether we can truly trust these explanations in high-stakes applications like healthcare or autonomous vehicles.

Current explanation methods lack standardized robustness evaluation: The paper identifies a critical gap in the field: while many methods exist to explain AI decisions, there is no consistent framework for measuring whether these explanations are actually reliable and resistant to manipulation, making it difficult for practitioners to choose trustworthy explanation methods.

Robustness of explanations is distinct from model accuracy: The survey emphasizes an important distinction—a model can make correct predictions while providing unreliable explanations of how it reached those decisions, meaning that verifying a model works well is not sufficient for ensuring responsible AI deployment.

Both attack and defense strategies for explanations are emerging: The research community is developing methods both to test explanation vulnerabilities (attacks) and to strengthen explanations against manipulation (defenses), though the survey suggests this defensive work is still in early stages.

Robust explanations are essential for safety-critical applications: As deep neural networks increasingly make decisions in risk-sensitive domains, the paper underscores that explanations must be robust and trustworthy—not just present—to meet the standards of responsible and accountable AI systems.